Multi-agent Coordination Algorithms for Pursuit-Evasion

Publication Type:
Thesis
Issue Date:
2023
Full metadata record
The real-world application of multi-agent systems (MAS) is still challenging and limited, despite the exclusive advantage of MAS over single agent systems. The multi-agent coordination or swarm intelligence is a paramount concern. Based on pursuit-evasion games, this thesis investigates the safe distributed implicit coordination of large-scale multi-agent systems, as some self-organizing systems in nature. Particularly, five research questions are answered in the thesis: (1) how to resolve collisions in the multi-agent environment; (2) how to be cooperative in a distributed way; (3) how to allocate the multi-agent task; (4) how to overcome the partial observation; and (5) how to get the arms race from the adversarial co-evolution. Accordingly, a safety-constrained multi-agent pursuit-evasion platform: MatrixWorld is proposed, and three coordination algorithms are designed, i.e., the cooperative coevolutionary particle swarm optimization (CCPSO-R) algorithm, the biquadratic assignment problem solver and parallel CCPSO-R (BiPCCR) algorithm, and the fuzzy self-organizing cooperative coevolution (FSC2) algorithm. Furthermore, the co-evolution mechanism, autocurricula, self-play, arms races, and adversarial learning are explored in the multi-agent setting by literature review and experimental verification. Safe coordination of thousands of agents is achieved in the experiments.
Please use this identifier to cite or link to this item: